mohamed-ahmed-58059/wikisql-text2sql

WikiSQL Text-to-SQL (execution-ready)

A cleaned, execution-ready repackaging of WikiSQL for text-to-SQL fine-tuning and execution-accuracy evaluation. Each example pairs a natural-language question with a gold SQL query over a single-table schema — and every split ships a real SQLite database so predicted SQL can be run and compared by result set (not string-matched).

Splits

Split	Examples
train	55,339
dev	8,263
test	15,519

Columns

Column	Type	Description
question	string	Natural-language question
table_name	string	Concise table name
columns	list[string]	Column names (snake_case)
column_types	list[string]	Per-column type: text or real
gold_sql	string	Gold SQL — the training target and the query to execute
db_id	string	Split name (the DB the example belongs to)
source	string	Origin dataset (wikisql)

Databases (for execution accuracy)

Each split has a SQLite DB at dbs/{split}.db with real table/column names, so both gold and predicted real-name SQL execute directly. Text columns are COLLATE U_NOCASE (Unicode-aware, casefold) and numeric columns are REAL.

⚠️ You must register the U_NOCASE collation before querying, or SQLite raises no such collation sequence: U_NOCASE. It's baked into the table schema.

import sqlite3
from datasets import load_dataset
from huggingface_hub import hf_hub_download

REPO = "mohamed-ahmed-58059/wikisql-text2sql"
ds = load_dataset(REPO, split="dev")
db = hf_hub_download(REPO, "dbs/dev.db", repo_type="dataset")

con = sqlite3.connect(db)
con.create_collation("U_NOCASE", lambda a, b: (a.casefold() > b.casefold()) - (a.casefold() < b.casefold()))

ex = ds[0]
print(ex["question"])
print(ex["gold_sql"])
print(con.execute(ex["gold_sql"]).fetchall())   # execute the gold query

How it was built

Derived from WikiSQL: table names made concise, columns normalized to snake_case, the WikiSQL logical forms decoded into clean real-name SQL, and a hardened SQLite DB built per split (NOCASE text, comma-stripped REAL numerics) so execution-accuracy comparison is robust to cosmetic differences. Full pipeline: github.com/mohamed-ahmed-58059/hf-ml-platform-text2sql.finetune.

Fine-tuned model

A QLoRA fine-tune of Llama-3.1-8B trained on this data reaches 0.915 execution accuracy on the dev split: Llama-3.1-8B-text2sql-wikisql.

License & attribution

Derived from WikiSQL (Salesforce), released under the BSD 3-Clause License. If you use this data, please cite the original work:

Victor Zhong, Caiming Xiong, Richard Socher. Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning. arXiv:1709.00103, 2017.